The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted Markdown data from README files to generate structured YAML metadata for model cards.
Ingestion points: The skill reads remote README.md content from Hugging Face repositories using the huggingface_hub library via the extract-readme command.
Boundary markers: The documentation mentions that the parser ignores code blocks and examples, but there are no specific prompt delimiters or instructions to ignore potential commands embedded within the evaluation table data itself.
Capability inventory: The agent is instructed to execute local scripts, perform network operations to external APIs, and push metadata changes or create Pull Requests on Hugging Face repositories.
Sanitization: While the skill normalizes names for matching, the content of benchmark metrics and values is directly interpolated into YAML model-index metadata without strict schema validation or filtering for embedded instructions.
[COMMAND_EXECUTION]: The skill instructions depend on the execution of multiple Python scripts located in a local scripts/ directory.
Evidence: Numerous usage examples (e.g., uv run scripts/evaluation_manager.py) and the 'Runtime availability' section in detailed-guide.md specify running these scripts via the uv tool.
Observation: The documentation explicitly states that no scripts are bundled within the skill package, requiring the user's environment to contain the necessary implementations. This relies on the presence of verified local code to perform actions.

hugging-face-evaluation